MueLu_Hierarchy_def.hpp

Go to the documentation of this file.

// @HEADER
//
// ***********************************************************************
//
//        MueLu: A package for multigrid based preconditioning
//                  Copyright 2012 Sandia Corporation
//
// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
// the U.S. Government retains certain rights in this software.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
// 1. Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright
// notice, this list of conditions and the following disclaimer in the
// documentation and/or other materials provided with the distribution.
//
// 3. Neither the name of the Corporation nor the names of the
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//
// Questions? Contact
//                    Jonathan Hu       (jhu@sandia.gov)
//                    Andrey Prokopenko (aprokop@sandia.gov)
//                    Ray Tuminaro      (rstumin@sandia.gov)
//
// ***********************************************************************
//
// @HEADER

#ifndef MUELU_HIERARCHY_DEF_HPP
#define MUELU_HIERARCHY_DEF_HPP

#include <time.h>

#include <algorithm>
#include <sstream>

#include <Xpetra_Matrix.hpp>
#include <Xpetra_MultiVectorFactory.hpp>
#include <Xpetra_Operator.hpp>
#include <Xpetra_IO.hpp>

#include "MueLu_Hierarchy_decl.hpp"

#include "MueLu_BoostGraphviz.hpp"
#include "MueLu_FactoryManager.hpp"
#include "MueLu_HierarchyUtils.hpp"
#include "MueLu_TopRAPFactory.hpp"
#include "MueLu_TopSmootherFactory.hpp"
#include "MueLu_Level.hpp"
#include "MueLu_Monitor.hpp"
#include "MueLu_PerfUtils.hpp"
#include "MueLu_PFactory.hpp"
#include "MueLu_SmootherFactoryBase.hpp"
#include "MueLu_SmootherFactory.hpp"
#include "MueLu_SmootherBase.hpp"

#include "Teuchos_TimeMonitor.hpp"



namespace MueLu {

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Hierarchy()
    : maxCoarseSize_(GetDefaultMaxCoarseSize()), implicitTranspose_(GetDefaultImplicitTranspose()),
      fuseProlongationAndUpdate_(GetDefaultFuseProlongationAndUpdate()),
      doPRrebalance_(GetDefaultPRrebalance()), isPreconditioner_(true), Cycle_(GetDefaultCycle()),
      scalingFactor_(Teuchos::ScalarTraits<double>::one()), lib_(Xpetra::UseTpetra), isDumpingEnabled_(false), dumpLevel_(-1), rate_(-1),
      sizeOfAllocatedLevelMultiVectors_(0)
  {
    AddLevel(rcp(new Level));
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Hierarchy(const std::string& label)
    : Hierarchy()
  {
    setObjectLabel(label);
    Levels_[0]->setObjectLabel(label);
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Hierarchy(const RCP<Matrix>& A)
    : maxCoarseSize_(GetDefaultMaxCoarseSize()), implicitTranspose_(GetDefaultImplicitTranspose()),
      fuseProlongationAndUpdate_(GetDefaultFuseProlongationAndUpdate()),
      doPRrebalance_(GetDefaultPRrebalance()), isPreconditioner_(true), Cycle_(GetDefaultCycle()),
      scalingFactor_(Teuchos::ScalarTraits<double>::one()), isDumpingEnabled_(false), dumpLevel_(-1), rate_(-1),
      sizeOfAllocatedLevelMultiVectors_(0)
  {
    lib_ = A->getDomainMap()->lib();

    RCP<Level> Finest = rcp(new Level);
    AddLevel(Finest);

    Finest->Set("A", A);
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Hierarchy(const RCP<Matrix>& A, const std::string& label)
    : Hierarchy(A)
  {
    setObjectLabel(label);
    Levels_[0]->setObjectLabel(label);
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::AddLevel(const RCP<Level>& level) {
    int levelID = LastLevelID() + 1; // ID of the inserted level

    if (level->GetLevelID() != -1 && (level->GetLevelID() != levelID))
      GetOStream(Warnings1) << "Hierarchy::AddLevel(): Level with ID=" << level->GetLevelID() <<
          " have been added at the end of the hierarchy\n but its ID have been redefined" <<
          " because last level ID of the hierarchy was " << LastLevelID() << "." << std::endl;

    Levels_.push_back(level);
    level->SetLevelID(levelID);
    level->setlib(lib_);

    level->SetPreviousLevel( (levelID == 0) ? Teuchos::null : Levels_[LastLevelID() - 1] );
    level->setObjectLabel(this->getObjectLabel());
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::AddNewLevel() {
    RCP<Level> newLevel = Levels_[LastLevelID()]->Build(); // new coarse level, using copy constructor
    newLevel->setlib(lib_);
    this->AddLevel(newLevel);                              // add to hierarchy
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  RCP<Level> & Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::GetLevel(const int levelID) {
    TEUCHOS_TEST_FOR_EXCEPTION(levelID < 0 || levelID > LastLevelID(), Exceptions::RuntimeError,
                               "MueLu::Hierarchy::GetLevel(): invalid input parameter value: LevelID = " << levelID);
    return Levels_[levelID];
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  int Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::GetNumLevels() const {
    return Levels_.size();
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  int Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::GetGlobalNumLevels() const {
    RCP<Operator> A = Levels_[0]->template Get<RCP<Operator> >("A");
    RCP<const Teuchos::Comm<int> > comm = A->getDomainMap()->getComm();

    int numLevels = GetNumLevels();
    int numGlobalLevels;
    Teuchos::reduceAll(*comm, Teuchos::REDUCE_MAX, numLevels, Teuchos::ptr(&numGlobalLevels));

    return numGlobalLevels;
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  double Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::GetOperatorComplexity() const {
    double totalNnz = 0, lev0Nnz = 1;
    for (int i = 0; i < GetNumLevels(); ++i) {
      TEUCHOS_TEST_FOR_EXCEPTION(!(Levels_[i]->IsAvailable("A")) , Exceptions::RuntimeError,
                                 "Operator complexity cannot be calculated because A is unavailable on level " << i);
      RCP<Operator> A = Levels_[i]->template Get<RCP<Operator> >("A");
      if (A.is_null())
        break;

      RCP<Matrix> Am = rcp_dynamic_cast<Matrix>(A);
      if (Am.is_null()) {
        GetOStream(Warnings0) << "Some level operators are not matrices, operator complexity calculation aborted" << std::endl;
        return 0.0;
      }

      totalNnz += as<double>(Am->getGlobalNumEntries());
      if (i == 0)
        lev0Nnz = totalNnz;
    }
    return totalNnz / lev0Nnz;
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  double Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::GetSmootherComplexity() const {
    double node_sc = 0, global_sc=0;
    double a0_nnz =0;
    const size_t INVALID = Teuchos::OrdinalTraits<size_t>::invalid();
    // Get cost of fine matvec
    if (GetNumLevels() <= 0) return -1.0;
    if (!Levels_[0]->IsAvailable("A")) return -1.0;

    RCP<Operator> A = Levels_[0]->template Get<RCP<Operator> >("A");
    if (A.is_null()) return -1.0;
    RCP<Matrix> Am = rcp_dynamic_cast<Matrix>(A);
    if(Am.is_null()) return -1.0;
    a0_nnz = as<double>(Am->getGlobalNumEntries());

    // Get smoother complexity at each level
    for (int i = 0; i < GetNumLevels(); ++i) {
      size_t level_sc=0;
      if(!Levels_[i]->IsAvailable("PreSmoother")) continue;
      RCP<SmootherBase> S = Levels_[i]->template Get<RCP<SmootherBase> >("PreSmoother");
      if (S.is_null()) continue;
      level_sc = S->getNodeSmootherComplexity();
      if(level_sc == INVALID) {global_sc=-1.0;break;}

      node_sc += as<double>(level_sc);
    }

    double min_sc=0.0;
    RCP<const Teuchos::Comm<int> > comm =A->getDomainMap()->getComm();
    Teuchos::reduceAll(*comm,Teuchos::REDUCE_SUM,node_sc,Teuchos::ptr(&global_sc));
    Teuchos::reduceAll(*comm,Teuchos::REDUCE_MIN,node_sc,Teuchos::ptr(&min_sc));

    if(min_sc < 0.0) return -1.0;
    else return global_sc / a0_nnz;
  }




  // Coherence checks todo in Setup() (using an helper function):
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::CheckLevel(Level& level, int levelID) {
    TEUCHOS_TEST_FOR_EXCEPTION(level.lib() != lib_, Exceptions::RuntimeError,
                               "MueLu::Hierarchy::CheckLevel(): wrong underlying linear algebra library.");
    TEUCHOS_TEST_FOR_EXCEPTION(level.GetLevelID() != levelID, Exceptions::RuntimeError,
                               "MueLu::Hierarchy::CheckLevel(): wrong level ID");
    TEUCHOS_TEST_FOR_EXCEPTION(levelID != 0 && level.GetPreviousLevel() != Levels_[levelID-1], Exceptions::RuntimeError,
                               "MueLu::Hierarchy::Setup(): wrong level parent");
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::SetMatvecParams(RCP<ParameterList> matvecParams) {
    for (int i = 0; i < GetNumLevels(); ++i) {
      RCP<Level> level = Levels_[i];
      if (level->IsAvailable("A")) {
        RCP<Operator> Aop = level->Get<RCP<Operator> >("A");
        RCP<Matrix> A = rcp_dynamic_cast<Matrix>(Aop);
        if (!A.is_null()) {
          RCP<const Import> xpImporter = A->getCrsGraph()->getImporter();
          if (!xpImporter.is_null())
            xpImporter->setDistributorParameters(matvecParams);
          RCP<const Export> xpExporter = A->getCrsGraph()->getExporter();
          if (!xpExporter.is_null())
            xpExporter->setDistributorParameters(matvecParams);
        }
      }
      if (level->IsAvailable("P")) {
        RCP<Matrix> P = level->Get<RCP<Matrix> >("P");
        RCP<const Import> xpImporter = P->getCrsGraph()->getImporter();
        if (!xpImporter.is_null())
          xpImporter->setDistributorParameters(matvecParams);
        RCP<const Export> xpExporter = P->getCrsGraph()->getExporter();
        if (!xpExporter.is_null())
          xpExporter->setDistributorParameters(matvecParams);
      }
      if (level->IsAvailable("R")) {
        RCP<Matrix> R = level->Get<RCP<Matrix> >("R");
        RCP<const Import> xpImporter = R->getCrsGraph()->getImporter();
        if (!xpImporter.is_null())
          xpImporter->setDistributorParameters(matvecParams);
        RCP<const Export> xpExporter = R->getCrsGraph()->getExporter();
        if (!xpExporter.is_null())
          xpExporter->setDistributorParameters(matvecParams);
      }
      if (level->IsAvailable("Importer")) {
        RCP<const Import> xpImporter = level->Get< RCP<const Import> >("Importer");
        if (!xpImporter.is_null())
          xpImporter->setDistributorParameters(matvecParams);
      }
    }
  }

  // The function uses three managers: fine, coarse and next coarse
  // We construct the data for the coarse level, and do requests for the next coarse
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Setup(int coarseLevelID,
                                                                   const RCP<const FactoryManagerBase> fineLevelManager,
                                                                   const RCP<const FactoryManagerBase> coarseLevelManager,
                                                                   const RCP<const FactoryManagerBase> nextLevelManager) {
    // Use PrintMonitor/TimerMonitor instead of just a FactoryMonitor to print "Level 0" instead of Hierarchy(0)
    // Print is done after the requests for next coarse level

    TEUCHOS_TEST_FOR_EXCEPTION(LastLevelID() < coarseLevelID, Exceptions::RuntimeError,
                               "MueLu::Hierarchy:Setup(): level " << coarseLevelID << " (specified by coarseLevelID argument) "
                               "must be built before calling this function.");

    Level& level = *Levels_[coarseLevelID];

    std::string label        = FormattingHelper::getColonLabel(level.getObjectLabel());
    TimeMonitor m1(*this, label + this->ShortClassName() + ": " + "Setup (total)");
    TimeMonitor m2(*this, label + this->ShortClassName() + ": " + "Setup" + " (total, level=" + Teuchos::toString(coarseLevelID) + ")");

    // TODO: pass coarseLevelManager by reference
    TEUCHOS_TEST_FOR_EXCEPTION(coarseLevelManager == Teuchos::null, Exceptions::RuntimeError,
                               "MueLu::Hierarchy::Setup(): argument coarseLevelManager cannot be null");

    typedef MueLu::TopRAPFactory<Scalar,LocalOrdinal,GlobalOrdinal,Node>      TopRAPFactory;
    typedef MueLu::TopSmootherFactory<Scalar,LocalOrdinal,GlobalOrdinal,Node> TopSmootherFactory;

    if (levelManagers_.size() < coarseLevelID+1)
      levelManagers_.resize(coarseLevelID+1);
    levelManagers_[coarseLevelID] = coarseLevelManager;

    bool isFinestLevel = (fineLevelManager.is_null());
    bool isLastLevel   = (nextLevelManager.is_null());

    int oldRank = -1;
    if (isFinestLevel) {
      RCP<Operator>                  A         = level.Get< RCP<Operator> >("A");
      RCP<const Map>                 domainMap = A->getDomainMap();
      RCP<const Teuchos::Comm<int> > comm      = domainMap->getComm();

      // Initialize random seed for reproducibility
      Utilities::SetRandomSeed(*comm);

      // Record the communicator on the level (used for timers sync)
      level.SetComm(comm);
      oldRank = SetProcRankVerbose(comm->getRank());

      // Set the Hierarchy library to match that of the finest level matrix,
      // even if it was already set
      lib_ = domainMap->lib();
      level.setlib(lib_);

    } else {
      // Permeate library to a coarser level
      level.setlib(lib_);

      Level& prevLevel = *Levels_[coarseLevelID-1];
      oldRank = SetProcRankVerbose(prevLevel.GetComm()->getRank());
    }

    CheckLevel(level, coarseLevelID);

    // Attach FactoryManager to the fine level
    RCP<SetFactoryManager> SFMFine;
    if (!isFinestLevel)
      SFMFine = rcp(new SetFactoryManager(Levels_[coarseLevelID-1], fineLevelManager));

    if (isFinestLevel && Levels_[coarseLevelID]->IsAvailable("Coordinates"))
      ReplaceCoordinateMap(*Levels_[coarseLevelID]);

    // Attach FactoryManager to the coarse level
    SetFactoryManager SFMCoarse(Levels_[coarseLevelID], coarseLevelManager);

    if (isDumpingEnabled_ && dumpLevel_ == 0 && coarseLevelID == 1)
      DumpCurrentGraph();

    RCP<TopSmootherFactory> coarseFact   = rcp(new TopSmootherFactory(coarseLevelManager, "CoarseSolver"));
    RCP<TopSmootherFactory> smootherFact = rcp(new TopSmootherFactory(coarseLevelManager, "Smoother"));

    int nextLevelID = coarseLevelID + 1;

    RCP<SetFactoryManager> SFMNext;
    if (isLastLevel == false) {
      // We are not at the coarsest level, so there is going to be another level ("next coarse") after this one ("coarse")
      if (nextLevelID > LastLevelID())
        AddNewLevel();
      CheckLevel(*Levels_[nextLevelID], nextLevelID);

      // Attach FactoryManager to the next level (level after coarse)
      SFMNext = rcp(new SetFactoryManager(Levels_[nextLevelID], nextLevelManager));
      Levels_[nextLevelID]->Request(TopRAPFactory(coarseLevelManager, nextLevelManager));

      // Do smoother requests here. We don't know whether this is going to be
      // the coarsest level or not, but we need to DeclareInput before we call
      // coarseRAPFactory.Build(), otherwise some stuff may be erased after
      // level releases
      level.Request(*smootherFact);

    } else {
      // Similar to smoother above, do the coarse solver request here. We don't
      // know whether this is going to be the coarsest level or not, but we
      // need to DeclareInput before we call coarseRAPFactory.Build(),
      // otherwise some stuff may be erased after level releases. This is
      // actually evident on ProjectorSmoother. It requires both "A" and
      // "Nullspace". However, "Nullspace" is erased after all releases, so if
      // we call the coarse factory request after RAP build we would not have
      // any data, and cannot get it as we don't have previous managers. The
      // typical trace looks like this:
      //
      // MueLu::Level(0)::GetFactory(Aggregates, 0): No FactoryManager
      //   during request for data "     Aggregates" on level 0 by factory TentativePFactory
      //   during request for data "              P" on level 1 by factory EminPFactory
      //   during request for data "              P" on level 1 by factory TransPFactory
      //   during request for data "              R" on level 1 by factory RAPFactory
      //   during request for data "              A" on level 1 by factory TentativePFactory
      //   during request for data "      Nullspace" on level 2 by factory NullspaceFactory
      //   during request for data "      Nullspace" on level 2 by factory NullspacePresmoothFactory
      //   during request for data "      Nullspace" on level 2 by factory ProjectorSmoother
      //   during request for data "    PreSmoother" on level 2 by factory NoFactory
      level.Request(*coarseFact);
    }

    PrintMonitor m0(*this, "Level " +  Teuchos::toString(coarseLevelID), static_cast<MsgType>(Runtime0 | Test));

    // Build coarse level hierarchy
    RCP<Operator> Ac = Teuchos::null;
    TopRAPFactory coarseRAPFactory(fineLevelManager, coarseLevelManager);

    if (level.IsAvailable("A")) {
      Ac = level.Get<RCP<Operator> >("A");
    } else if (!isFinestLevel) {
      // We only build here, the release is done later
      coarseRAPFactory.Build(*level.GetPreviousLevel(), level);
    }

    if (level.IsAvailable("A"))
      Ac = level.Get<RCP<Operator> >("A");
    RCP<Matrix> Acm = rcp_dynamic_cast<Matrix>(Ac);

    // Record the communicator on the level
    if (!Ac.is_null())
      level.SetComm(Ac->getDomainMap()->getComm());

    // Test if we reach the end of the hierarchy
    bool isOrigLastLevel = isLastLevel;
    if (isLastLevel) {
      // Last level as we have achieved the max limit
      isLastLevel = true;

    } else if (Ac.is_null()) {
      // Last level for this processor, as it does not belong to the next
      // subcommunicator. Other processors may continue working on the
      // hierarchy
      isLastLevel = true;

    } else {
      if (!Acm.is_null() && Acm->getGlobalNumRows() <= maxCoarseSize_) {
        // Last level as the size of the coarse matrix became too small
        GetOStream(Runtime0) << "Max coarse size (<= " << maxCoarseSize_ << ") achieved" << std::endl;
        isLastLevel = true;
      }
    }

    if (!Ac.is_null() && !isFinestLevel) {
      RCP<Operator> A  = Levels_[coarseLevelID-1]->template Get< RCP<Operator> >("A");
      RCP<Matrix>   Am = rcp_dynamic_cast<Matrix>(A);

      const double maxCoarse2FineRatio = 0.8;
      if (!Acm.is_null() && !Am.is_null() && Acm->getGlobalNumRows() > maxCoarse2FineRatio * Am->getGlobalNumRows()) {
        // We could abort here, but for now we simply notify user.
        // Couple of additional points:
        //   - if repartitioning is delayed until level K, but the aggregation
        //     procedure stagnates between levels K-1 and K.  In this case,
        //     repartitioning could enable faster coarsening once again, but the
        //     hierarchy construction will abort due to the stagnation check.
        //   - if the matrix is small enough, we could move it to one processor.
        GetOStream(Warnings0) << "Aggregation stagnated. Please check your matrix and/or adjust your configuration file."
            << "Possible fixes:\n"
            << "  - reduce the maximum number of levels\n"
            << "  - enable repartitioning\n"
            << "  - increase the minimum coarse size." << std::endl;

      }
    }

    if (isLastLevel) {
      if (!isOrigLastLevel) {
        // We did not expect to finish this early so we did request a smoother.
        // We need a coarse solver instead. Do the magic.
        level.Release(*smootherFact);
        level.Request(*coarseFact);
      }

      // Do the actual build, if we have any data.
      // NOTE: this is not a great check, we may want to call Build() regardless.
      if (!Ac.is_null())
        coarseFact->Build(level);

      // Once the dirty deed is done, release stuff. The smoother has already
      // been released.
      level.Release(*coarseFact);

    } else {
      // isLastLevel = false => isOrigLastLevel = false, meaning that we have
      // requested the smoother. Now we need to build it and to release it.
      // We don't need to worry about the coarse solver, as we didn't request it.
      if (!Ac.is_null())
        smootherFact->Build(level);

      level.Release(*smootherFact);
    }

    if (isLastLevel == true) {
      if (isOrigLastLevel == false) {
        // Earlier in the function, we constructed the next coarse level, and requested data for the that level,
        // assuming that we are not at the coarsest level. Now, we changed our mind, so we have to release those.
        Levels_[nextLevelID]->Release(TopRAPFactory(coarseLevelManager, nextLevelManager));
      }
      Levels_.resize(nextLevelID);
    }

    // I think this is the proper place for graph so that it shows every dependence
    if (isDumpingEnabled_ && dumpLevel_ > 0 && coarseLevelID == dumpLevel_)
      DumpCurrentGraph();

    if (!isFinestLevel) {
      // Release the hierarchy data
      // We release so late to help blocked solvers, as the smoothers for them need A blocks
      // which we construct in RAPFactory
      level.Release(coarseRAPFactory);
    }

    if (oldRank != -1)
      SetProcRankVerbose(oldRank);

    return isLastLevel;
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::SetupRe() {
    int numLevels = Levels_.size();
    TEUCHOS_TEST_FOR_EXCEPTION(levelManagers_.size() != numLevels, Exceptions::RuntimeError,
                               "Hierarchy::SetupRe: " << Levels_.size() << " levels, but  " << levelManagers_.size() << " level factory managers");

    const int startLevel = 0;
    Clear(startLevel);

#ifdef HAVE_MUELU_DEBUG
      // Reset factories' data used for debugging
      for (int i = 0; i < numLevels; i++)
        levelManagers_[i]->ResetDebugData();

#endif

    int levelID;
    for (levelID = startLevel; levelID < numLevels;) {
      bool r = Setup(levelID,
                     (levelID != 0 ? levelManagers_[levelID-1] : Teuchos::null),
                     levelManagers_[levelID],
                     (levelID+1 != numLevels ? levelManagers_[levelID+1] : Teuchos::null));
      levelID++;
      if (r) break;
    }
    // We may construct fewer levels for some reason, make sure we continue
    // doing that in the future
    Levels_       .resize(levelID);
    levelManagers_.resize(levelID);

    // NOTE: All reuse cases leave all of the maps the same, meaning that we do not
    // need to reallocated the cached multivectors for Iterate().  If this were to change,
    // we'd want to do a DeleteLevelMultiVectors() and AllocateLevelMultiVectors() here.

    // since the # of levels, etc. may have changed, force re-determination of description during next call to description()
    ResetDescription();

    describe(GetOStream(Statistics0), GetVerbLevel());
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Setup(const FactoryManagerBase& manager, int startLevel, int numDesiredLevels) {
    // Use MueLu::BaseClass::description() to avoid printing "{numLevels = 1}" (numLevels is increasing...)
    PrintMonitor m0(*this, "Setup (" + this->MueLu::BaseClass::description() + ")", Runtime0);

    Clear(startLevel);

    // Check Levels_[startLevel] exists.
    TEUCHOS_TEST_FOR_EXCEPTION(Levels_.size() <= startLevel, Exceptions::RuntimeError,
                               "MueLu::Hierarchy::Setup(): fine level (" << startLevel << ") does not exist");

    TEUCHOS_TEST_FOR_EXCEPTION(numDesiredLevels <= 0, Exceptions::RuntimeError,
                               "Constructing non-positive (" << numDesiredLevels << ") number of levels does not make sense.");

    // Check for fine level matrix A
    TEUCHOS_TEST_FOR_EXCEPTION(!Levels_[startLevel]->IsAvailable("A"), Exceptions::RuntimeError,
                               "MueLu::Hierarchy::Setup(): fine level (" << startLevel << ") has no matrix A! "
                               "Set fine level matrix A using Level.Set()");

    RCP<Operator> A = Levels_[startLevel]->template Get<RCP<Operator> >("A");
    lib_ = A->getDomainMap()->lib();

    if (IsPrint(Statistics2)) {
      RCP<Matrix> Amat = rcp_dynamic_cast<Matrix>(A);

      if (!Amat.is_null()) {
          RCP<ParameterList> params = rcp(new ParameterList());
          params->set("printLoadBalancingInfo", true);
          params->set("printCommInfo",          true);

          GetOStream(Statistics2) << PerfUtils::PrintMatrixInfo(*Amat, "A0", params);
      } else {
          GetOStream(Warnings1) << "Fine level operator is not a matrix, statistics are not available" << std::endl;
      }
    }

    RCP<const FactoryManagerBase> rcpmanager = rcpFromRef(manager);

    const int lastLevel = startLevel + numDesiredLevels - 1;
    GetOStream(Runtime0) << "Setup loop: startLevel = " << startLevel << ", lastLevel = " << lastLevel
        << " (stop if numLevels = " << numDesiredLevels << " or Ac.size() < " << maxCoarseSize_ << ")" << std::endl;

    // Setup multigrid levels
    int iLevel = 0;
    if (numDesiredLevels == 1) {
      iLevel = 0;
      Setup(startLevel, Teuchos::null, rcpmanager, Teuchos::null);                     // setup finest==coarsest level (first and last managers are Teuchos::null)

    } else {
      bool bIsLastLevel = Setup(startLevel, Teuchos::null, rcpmanager, rcpmanager);    // setup finest level (level 0) (first manager is Teuchos::null)
      if (bIsLastLevel == false) {
        for (iLevel = startLevel + 1; iLevel < lastLevel; iLevel++) {
          bIsLastLevel = Setup(iLevel, rcpmanager, rcpmanager, rcpmanager);            // setup intermediate levels
          if (bIsLastLevel == true)
            break;
        }
        if (bIsLastLevel == false)
          Setup(lastLevel, rcpmanager, rcpmanager, Teuchos::null);                     // setup coarsest level (last manager is Teuchos::null)
      }
    }

    // TODO: some check like this should be done at the beginning of the routine
    TEUCHOS_TEST_FOR_EXCEPTION(iLevel != Levels_.size() - 1, Exceptions::RuntimeError,
                               "MueLu::Hierarchy::Setup(): number of level");

    // TODO: this is not exception safe: manager will still hold default
    // factories if you exit this function with an exception
    manager.Clean();

    describe(GetOStream(Statistics0), GetVerbLevel());
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Clear(int startLevel) {
    if (startLevel < GetNumLevels())
      GetOStream(Runtime0) << "Clearing old data (if any)" << std::endl;

    for (int iLevel = startLevel; iLevel < GetNumLevels(); iLevel++)
      Levels_[iLevel]->Clear();
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::ExpertClear() {
    GetOStream(Runtime0) << "Clearing old data (expert)" << std::endl;
    for (int iLevel = 0; iLevel < GetNumLevels(); iLevel++)
      Levels_[iLevel]->ExpertClear();
  }

#if defined(HAVE_MUELU_EXPERIMENTAL) && defined(HAVE_MUELU_ADDITIVE_VARIANT)
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  ReturnType Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Iterate(const MultiVector& B, MultiVector& X, ConvData conv,
                                                                           bool InitialGuessIsZero, LO startLevel) {
    LO            nIts = conv.maxIts_;
    MagnitudeType tol  = conv.tol_;

    std::string prefix       = this->ShortClassName() + ": ";
    std::string levelSuffix  = " (level=" + toString(startLevel) + ")";
    std::string levelSuffix1 = " (level=" + toString(startLevel+1) + ")";

    using namespace Teuchos;
    RCP<Time> CompTime                 = Teuchos::TimeMonitor::getNewCounter(prefix + "Computational Time (total)");
    RCP<Time> Concurrent               = Teuchos::TimeMonitor::getNewCounter(prefix + "Concurrent portion");
    RCP<Time> ApplyR                   = Teuchos::TimeMonitor::getNewCounter(prefix + "R: Computational Time");
    RCP<Time> ApplyPbar                = Teuchos::TimeMonitor::getNewCounter(prefix + "Pbar: Computational Time");
    RCP<Time> CompFine                 = Teuchos::TimeMonitor::getNewCounter(prefix + "Fine: Computational Time");
    RCP<Time> CompCoarse               = Teuchos::TimeMonitor::getNewCounter(prefix + "Coarse: Computational Time");
    RCP<Time> ApplySum                 = Teuchos::TimeMonitor::getNewCounter(prefix + "Sum: Computational Time");
    RCP<Time> Synchronize_beginning    = Teuchos::TimeMonitor::getNewCounter(prefix + "Synchronize_beginning");
    RCP<Time> Synchronize_center       = Teuchos::TimeMonitor::getNewCounter(prefix + "Synchronize_center");
    RCP<Time> Synchronize_end          = Teuchos::TimeMonitor::getNewCounter(prefix + "Synchronize_end");

    RCP<Level>          Fine = Levels_[0];
    RCP<Level>          Coarse;

    RCP<Operator> A = Fine->Get< RCP<Operator> >("A");
    Teuchos::RCP< const Teuchos::Comm< int > > communicator = A->getDomainMap()->getComm();


    //Synchronize_beginning->start();
    //communicator->barrier();
    //Synchronize_beginning->stop();

    CompTime->start();

    SC one = STS::one(), zero = STS::zero();

    bool zeroGuess = InitialGuessIsZero;

    // =======   UPFRONT DEFINITION OF COARSE VARIABLES ===========

    //RCP<const Map> origMap;
    RCP< Operator > P;
    RCP< Operator > Pbar;
    RCP< Operator > R;
    RCP< MultiVector > coarseRhs, coarseX;
    RCP< Operator > Ac;
    RCP<SmootherBase> preSmoo_coarse, postSmoo_coarse;
    bool emptyCoarseSolve = true;
    RCP<MultiVector> coarseX_prolonged = MultiVectorFactory::Build(X.getMap(), X.getNumVectors(), true);

    RCP<const Import> importer;

    if (Levels_.size() > 1) {
      Coarse = Levels_[1];
      if (Coarse->IsAvailable("Importer"))
        importer = Coarse->Get< RCP<const Import> >("Importer");

      R = Coarse->Get< RCP<Operator> >("R");
      P = Coarse->Get< RCP<Operator> >("P");


      //if(Coarse->IsAvailable("Pbar"))
      Pbar = Coarse->Get< RCP<Operator> >("Pbar");

      coarseRhs = MultiVectorFactory::Build(R->getRangeMap(), B.getNumVectors(), true);

      Ac = Coarse->Get< RCP< Operator > >("A");

      ApplyR->start();
      R->apply(B, *coarseRhs, Teuchos::NO_TRANS, one, zero);
      //P->apply(B, *coarseRhs, Teuchos::TRANS, one, zero);
      ApplyR->stop();

      if (doPRrebalance_ || importer.is_null()) {
        coarseX = MultiVectorFactory::Build(coarseRhs->getMap(), X.getNumVectors(), true);

      } else {

        RCP<TimeMonitor> ITime      = rcp(new TimeMonitor(*this, prefix + "Solve : import (total)"       , Timings0));
        RCP<TimeMonitor> ILevelTime = rcp(new TimeMonitor(*this, prefix + "Solve : import" + levelSuffix1, Timings0));

        // Import: range map of R --> domain map of rebalanced Ac (before subcomm replacement)
        RCP<MultiVector> coarseTmp = MultiVectorFactory::Build(importer->getTargetMap(), coarseRhs->getNumVectors());
        coarseTmp->doImport(*coarseRhs, *importer, Xpetra::INSERT);
        coarseRhs.swap(coarseTmp);

        coarseX = MultiVectorFactory::Build(importer->getTargetMap(), X.getNumVectors(), true);
      }

      if (Coarse->IsAvailable("PreSmoother"))
        preSmoo_coarse = Coarse->Get< RCP<SmootherBase> >("PreSmoother");
      if (Coarse->IsAvailable("PostSmoother"))
        postSmoo_coarse = Coarse->Get< RCP<SmootherBase> >("PostSmoother");
    }

    // ==========================================================


    MagnitudeType prevNorm = STS::magnitude(STS::one()), curNorm = STS::magnitude(STS::one());
    rate_ = 1.0;

    for (LO i = 1; i <= nIts; i++) {
#ifdef HAVE_MUELU_DEBUG
      if (A->getDomainMap()->isCompatible(*(X.getMap())) == false) {
        std::ostringstream ss;
        ss << "Level " << startLevel << ": level A's domain map is not compatible with X";
        throw Exceptions::Incompatible(ss.str());
      }

      if (A->getRangeMap()->isCompatible(*(B.getMap())) == false) {
        std::ostringstream ss;
        ss << "Level " << startLevel << ": level A's range map is not compatible with B";
        throw Exceptions::Incompatible(ss.str());
      }
#endif
    }

   bool emptyFineSolve = true;

   RCP< MultiVector > fineX;
   fineX = MultiVectorFactory::Build(X.getMap(), X.getNumVectors(), true);

   //Synchronize_center->start();
   //communicator->barrier();
   //Synchronize_center->stop();

   Concurrent->start();

   // NOTE: we need to check using IsAvailable before Get here to avoid building default smoother
   if (Fine->IsAvailable("PreSmoother")) {
     RCP<SmootherBase> preSmoo = Fine->Get< RCP<SmootherBase> >("PreSmoother");
     CompFine->start();
     preSmoo->Apply(*fineX, B, zeroGuess);
     CompFine->stop();
     emptyFineSolve = false;
   }
   if (Fine->IsAvailable("PostSmoother")) {
     RCP<SmootherBase> postSmoo = Fine->Get< RCP<SmootherBase> >("PostSmoother");
     CompFine->start();
     postSmoo->Apply(*fineX, B, zeroGuess);
     CompFine->stop();

     emptyFineSolve = false;
   }
   if (emptyFineSolve == true) {
     GetOStream(Warnings1) << "No fine grid smoother" << std::endl;
     // Fine grid smoother is identity
     fineX->update(one, B, zero);
   }

   if (Levels_.size() > 1) {
     // NOTE: we need to check using IsAvailable before Get here to avoid building default smoother
     if (Coarse->IsAvailable("PreSmoother")) {
       CompCoarse->start();
       preSmoo_coarse->Apply(*coarseX, *coarseRhs, zeroGuess);
       CompCoarse->stop();
       emptyCoarseSolve = false;

     }
     if (Coarse->IsAvailable("PostSmoother")) {
       CompCoarse->start();
       postSmoo_coarse->Apply(*coarseX, *coarseRhs, zeroGuess);
       CompCoarse->stop();
       emptyCoarseSolve = false;

     }
     if (emptyCoarseSolve == true) {
       GetOStream(Warnings1) << "No coarse grid solver" << std::endl;
       // Coarse operator is identity
       coarseX->update(one, *coarseRhs, zero);
     }
     Concurrent->stop();
     //Synchronize_end->start();
     //communicator->barrier();
     //Synchronize_end->stop();

     if (!doPRrebalance_ && !importer.is_null()) {
       RCP<TimeMonitor> ITime      = rcp(new TimeMonitor(*this, prefix + "Solve : export (total)"       , Timings0));
       RCP<TimeMonitor> ILevelTime = rcp(new TimeMonitor(*this, prefix + "Solve : export" + levelSuffix1, Timings0));

       // Import: range map of rebalanced Ac (before subcomm replacement) --> domain map of P
       RCP<MultiVector> coarseTmp = MultiVectorFactory::Build(importer->getSourceMap(), coarseX->getNumVectors());
       coarseTmp->doExport(*coarseX, *importer, Xpetra::INSERT);
       coarseX.swap(coarseTmp);
     }

     ApplyPbar->start();
     Pbar->apply(*coarseX, *coarseX_prolonged, Teuchos::NO_TRANS, one, zero);
     ApplyPbar->stop();
   }

   ApplySum->start();
   X.update(1.0, *fineX, 1.0, *coarseX_prolonged, 0.0);
   ApplySum->stop();

   CompTime->stop();

   //communicator->barrier();

   return (tol > 0 ? Unconverged : Undefined);
}
#else
  // ---------------------------------------- Iterate -------------------------------------------------------
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  ReturnType Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Iterate(const MultiVector& B, MultiVector& X, ConvData conv,
                                                                           bool InitialGuessIsZero, LO startLevel) {
    LO            nIts = conv.maxIts_;
    MagnitudeType tol  = conv.tol_;

    // These timers work as follows. "iterateTime" records total time spent in
    // iterate. "levelTime" records time on a per level basis. The label is
    // crafted to mimic the per-level messages used in Monitors. Note that a
    // given level is timed with a TimeMonitor instead of a Monitor or
    // SubMonitor. This is mainly because I want to time each level
    // separately, and Monitors/SubMonitors print "(total) xx yy zz" ,
    // "(sub,total) xx yy zz", respectively, which is subject to
    // misinterpretation.  The per-level TimeMonitors are stopped/started
    // manually before/after a recursive call to Iterate. A side artifact to
    // this approach is that the counts for intermediate level timers are twice
    // the counts for the finest and coarsest levels.

    RCP<Level>    Fine = Levels_[startLevel];

    std::string label        = FormattingHelper::getColonLabel(Fine->getObjectLabel());
    std::string prefix       = label + this->ShortClassName() + ": ";
    std::string levelSuffix  = " (level=" + toString(startLevel) + ")";
    std::string levelSuffix1 = " (level=" + toString(startLevel+1) + ")";

    bool useStackedTimer = !Teuchos::TimeMonitor::getStackedTimer().is_null();

    RCP<Monitor>     iterateTime;
    RCP<TimeMonitor> iterateTime1;
    if (startLevel == 0)
      iterateTime  = rcp(new Monitor(*this, "Solve", label, (nIts == 1) ? None : Runtime0, Timings0));
    else if (!useStackedTimer)
      iterateTime1 = rcp(new TimeMonitor(*this, prefix + "Solve (total, level=" + toString(startLevel) + ")", Timings0));

    std::string iterateLevelTimeLabel = prefix + "Solve" + levelSuffix;
    RCP<TimeMonitor> iterateLevelTime = rcp(new TimeMonitor(*this, iterateLevelTimeLabel, Timings0));

    bool zeroGuess = InitialGuessIsZero;

    RCP<Operator> A    = Fine->Get< RCP<Operator> >("A");
    using namespace Teuchos;
    RCP<Time> CompCoarse  = Teuchos::TimeMonitor::getNewCounter(prefix + "Coarse: Computational Time");

    if (A.is_null()) {
      // This processor does not have any data for this process on coarser
      // levels. This can only happen when there are multiple processors and
      // we use repartitioning.
      return Undefined;
    }

    // If we switched the number of vectors, we'd need to reallocate here.
    // If the number of vectors is unchanged, this is a noop.
    // NOTE: We need to check against B because the tests in AllocateLevelMultiVectors
    // will fail on Stokhos Scalar types (due to the so-called 'hidden dimension')
    const BlockedMultiVector * Bblocked = dynamic_cast<const BlockedMultiVector*>(&B);  
    if(residual_.size() > startLevel &&
       ( ( Bblocked && !Bblocked->isSameSize(*residual_[startLevel])) ||
         (!Bblocked && !residual_[startLevel]->isSameSize(B))))
      DeleteLevelMultiVectors();
    AllocateLevelMultiVectors(X.getNumVectors());

    // Print residual information before iterating
    typedef Teuchos::ScalarTraits<typename STS::magnitudeType> STM;
    MagnitudeType prevNorm = STM::one(), curNorm = STM::one();
    rate_ = 1.0;
    if (startLevel == 0 && !isPreconditioner_ &&
        (IsPrint(Statistics1) || tol > 0)) {
      // We calculate the residual only if we want to print it out, or if we
      // want to stop once we achive the tolerance
      Teuchos::Array<MagnitudeType> rn;
      rn = Utilities::ResidualNorm(*A, X, B,*residual_[startLevel]);

      if (tol > 0) {
        bool passed = true;
        for (LO k = 0; k < rn.size(); k++)
          if (rn[k] >= tol)
            passed = false;

        if (passed)
          return Converged;
      }

      if (IsPrint(Statistics1))
        GetOStream(Statistics1) << "iter:    "
            << std::setiosflags(std::ios::left)
            << std::setprecision(3) << 0 // iter 0
            << "           residual = "
            << std::setprecision(10) << rn
            << std::endl;
    }

    SC one = STS::one(), zero = STS::zero();
    for (LO i = 1; i <= nIts; i++) {
#ifdef HAVE_MUELU_DEBUG
#if 0 // TODO fix me
      if (A->getDomainMap()->isCompatible(*(X.getMap())) == false) {
        std::ostringstream ss;
        ss << "Level " << startLevel << ": level A's domain map is not compatible with X";
        throw Exceptions::Incompatible(ss.str());
      }

      if (A->getRangeMap()->isCompatible(*(B.getMap())) == false) {
        std::ostringstream ss;
        ss << "Level " << startLevel << ": level A's range map is not compatible with B";
        throw Exceptions::Incompatible(ss.str());
      }
#endif
#endif

      if (startLevel == as<LO>(Levels_.size())-1) {
        // On the coarsest level, we do either smoothing (if defined) or a direct solve.
        RCP<TimeMonitor> CLevelTime = rcp(new TimeMonitor(*this, prefix + "Solve : coarse" + levelSuffix, Timings0));

        bool emptySolve = true;

        // NOTE: we need to check using IsAvailable before Get here to avoid building default smoother
        if (Fine->IsAvailable("PreSmoother")) {
          RCP<SmootherBase> preSmoo = Fine->Get< RCP<SmootherBase> >("PreSmoother");
          CompCoarse->start();
          preSmoo->Apply(X, B, zeroGuess);
          CompCoarse->stop();
          zeroGuess  = false;
          emptySolve = false;
        }
        if (Fine->IsAvailable("PostSmoother")) {
          RCP<SmootherBase> postSmoo = Fine->Get< RCP<SmootherBase> >("PostSmoother");
          CompCoarse->start();
          postSmoo->Apply(X, B, zeroGuess);
          CompCoarse->stop();
          emptySolve = false;
        }
        if (emptySolve == true) {
          GetOStream(Warnings1) << "No coarse grid solver" << std::endl;
          // Coarse operator is identity
          X.update(one, B, zero);
        }

      } else {
        // On intermediate levels, we do cycles
        RCP<Level> Coarse = Levels_[startLevel+1];
        {
          // ============== PRESMOOTHING ==============
          RCP<TimeMonitor> STime;
          if (!useStackedTimer)
            STime                     = rcp(new TimeMonitor(*this, prefix + "Solve : smoothing (total)"      , Timings0));
          RCP<TimeMonitor> SLevelTime = rcp(new TimeMonitor(*this, prefix + "Solve : smoothing" + levelSuffix, Timings0));

          if (Fine->IsAvailable("PreSmoother")) {
            RCP<SmootherBase> preSmoo = Fine->Get< RCP<SmootherBase> >("PreSmoother");
            preSmoo->Apply(X, B, zeroGuess);
          }
        }

        RCP<MultiVector> residual;
        {
          RCP<TimeMonitor> ATime;
          if (!useStackedTimer)
            ATime                     = rcp(new TimeMonitor(*this, prefix + "Solve : residual calculation (total)"      , Timings0));
          RCP<TimeMonitor> ALevelTime = rcp(new TimeMonitor(*this, prefix + "Solve : residual calculation" + levelSuffix, Timings0));
          Utilities::Residual(*A, X, B,*residual_[startLevel]);
          residual = residual_[startLevel];
        }

        RCP<Operator>    P = Coarse->Get< RCP<Operator> >("P");
        if (Coarse->IsAvailable("Pbar"))
           P = Coarse->Get< RCP<Operator> >("Pbar");

        RCP<MultiVector> coarseRhs, coarseX;
        //        const bool initializeWithZeros = true;
        {
          // ============== RESTRICTION ==============
          RCP<TimeMonitor> RTime;
          if (!useStackedTimer)
            RTime                     = rcp(new TimeMonitor(*this, prefix + "Solve : restriction (total)"      , Timings0));
          RCP<TimeMonitor> RLevelTime = rcp(new TimeMonitor(*this, prefix + "Solve : restriction" + levelSuffix, Timings0));
          coarseRhs = coarseRhs_[startLevel];

          if (implicitTranspose_) {
            P->apply(*residual, *coarseRhs, Teuchos::TRANS, one, zero);

          } else {
            RCP<Operator> R = Coarse->Get< RCP<Operator> >("R");
            R->apply(*residual, *coarseRhs, Teuchos::NO_TRANS, one, zero);
          }
        }

        RCP<const Import> importer;
        if (Coarse->IsAvailable("Importer"))
          importer = Coarse->Get< RCP<const Import> >("Importer");

        coarseX = coarseX_[startLevel];
        if (!doPRrebalance_ && !importer.is_null()) {
          RCP<TimeMonitor> ITime;
          if (!useStackedTimer)
            ITime                     = rcp(new TimeMonitor(*this, prefix + "Solve : import (total)"       , Timings0));
          RCP<TimeMonitor> ILevelTime = rcp(new TimeMonitor(*this, prefix + "Solve : import" + levelSuffix1, Timings0));

          // Import: range map of R --> domain map of rebalanced Ac (before subcomm replacement)
          RCP<MultiVector> coarseTmp = coarseImport_[startLevel];
          coarseTmp->doImport(*coarseRhs, *importer, Xpetra::INSERT);
          coarseRhs.swap(coarseTmp);
        }

        RCP<Operator> Ac = Coarse->Get< RCP<Operator> >("A");
        if (!Ac.is_null()) {
          RCP<const Map> origXMap = coarseX->getMap();
          RCP<const Map> origRhsMap = coarseRhs->getMap();

          // Replace maps with maps with a subcommunicator
          coarseRhs->replaceMap(Ac->getRangeMap());
          coarseX  ->replaceMap(Ac->getDomainMap());

          {
            iterateLevelTime = Teuchos::null; // stop timing this level

            Iterate(*coarseRhs, *coarseX, 1, true, startLevel+1);
            // ^^ zero initial guess
            if (Cycle_ == WCYCLE)
              Iterate(*coarseRhs, *coarseX, 1, false, startLevel+1);
            // ^^ nonzero initial guess

            iterateLevelTime = rcp(new TimeMonitor(*this, iterateLevelTimeLabel));  // restart timing this level
          }
          coarseX->replaceMap(origXMap);
          coarseRhs->replaceMap(origRhsMap);
        }

        if (!doPRrebalance_ && !importer.is_null()) {
          RCP<TimeMonitor> ITime;
          if (!useStackedTimer)
            ITime                     = rcp(new TimeMonitor(*this, prefix + "Solve : export (total)"      ,  Timings0));
          RCP<TimeMonitor> ILevelTime = rcp(new TimeMonitor(*this, prefix + "Solve : export" + levelSuffix1, Timings0));

          // Import: range map of rebalanced Ac (before subcomm replacement) --> domain map of P
          RCP<MultiVector> coarseTmp = coarseExport_[startLevel];
          coarseTmp->doExport(*coarseX, *importer, Xpetra::INSERT);
          coarseX.swap(coarseTmp);
        }

        {
          // ============== PROLONGATION ==============
          RCP<TimeMonitor> PTime;
          if (!useStackedTimer)
            PTime                     = rcp(new TimeMonitor(*this, prefix + "Solve : prolongation (total)"      , Timings0));
          RCP<TimeMonitor> PLevelTime = rcp(new TimeMonitor(*this, prefix + "Solve : prolongation" + levelSuffix, Timings0));
          // Update X += P * coarseX
          // Note that due to what may be round-off error accumulation, use of the fused kernel
          //    P->apply(*coarseX, X, Teuchos::NO_TRANS, one, one);
          // can in some cases result in slightly higher iteration counts.
          if (fuseProlongationAndUpdate_) {
            P->apply(*coarseX, X, Teuchos::NO_TRANS, scalingFactor_, one);
          } else {
            RCP<MultiVector> correction = correction_[startLevel];
            P->apply(*coarseX, *correction, Teuchos::NO_TRANS, one, zero);
            X.update(scalingFactor_, *correction, one);
          }
        }

        {
          // ============== POSTSMOOTHING ==============
          RCP<TimeMonitor> STime;
          if (!useStackedTimer)
            STime                     = rcp(new TimeMonitor(*this, prefix + "Solve : smoothing (total)"      , Timings0));
          RCP<TimeMonitor> SLevelTime = rcp(new TimeMonitor(*this, prefix + "Solve : smoothing" + levelSuffix, Timings0));

          if (Fine->IsAvailable("PostSmoother")) {
            RCP<SmootherBase> postSmoo = Fine->Get< RCP<SmootherBase> >("PostSmoother");
            postSmoo->Apply(X, B, false);
          }
        }
      }
      zeroGuess = false;

      if (startLevel == 0 && !isPreconditioner_ &&
          (IsPrint(Statistics1) || tol > 0)) {
        // We calculate the residual only if we want to print it out, or if we
        // want to stop once we achive the tolerance
        Teuchos::Array<MagnitudeType> rn;
        rn = Utilities::ResidualNorm(*A, X, B,*residual_[startLevel]);

        prevNorm = curNorm;
        curNorm  = rn[0];
        rate_ = as<MagnitudeType>(curNorm / prevNorm);

        if (IsPrint(Statistics1))
          GetOStream(Statistics1) << "iter:    "
                                     << std::setiosflags(std::ios::left)
                                     << std::setprecision(3) << i
                                     << "           residual = "
                                     << std::setprecision(10) << rn
                                     << std::endl;

        if (tol > 0) {
          bool passed = true;
          for (LO k = 0; k < rn.size(); k++)
            if (rn[k] >= tol)
              passed = false;

          if (passed)
            return Converged;
        }
      }
    }
    return (tol > 0 ? Unconverged : Undefined);
  }
#endif


  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Write(const LO& start, const LO& end, const std::string &suffix) {
    LO startLevel = (start != -1 ? start : 0);
    LO   endLevel = (end   != -1 ? end   : Levels_.size()-1);

    TEUCHOS_TEST_FOR_EXCEPTION(startLevel > endLevel, Exceptions::RuntimeError,
                               "MueLu::Hierarchy::Write : startLevel must be <= endLevel");

    TEUCHOS_TEST_FOR_EXCEPTION(startLevel < 0 || endLevel >= Levels_.size(), Exceptions::RuntimeError,
                               "MueLu::Hierarchy::Write bad start or end level");

    for (LO i = startLevel; i < endLevel + 1; i++) {
      RCP<Matrix> A = rcp_dynamic_cast<Matrix>(Levels_[i]-> template Get< RCP< Operator> >("A")), P, R;
      if (i > 0) {
        P = rcp_dynamic_cast<Matrix>(Levels_[i]-> template Get< RCP< Operator> >("P"));
        if (!implicitTranspose_)
          R = rcp_dynamic_cast<Matrix>(Levels_[i]-> template Get< RCP< Operator> >("R"));
      }

      if (!A.is_null()) Xpetra::IO<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Write("A_" + toString(i) + suffix + ".m", *A);
      if (!P.is_null()) {
        Xpetra::IO<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Write("P_" + toString(i) + suffix + ".m", *P);
      }
      if (!R.is_null()) {
        Xpetra::IO<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Write("R_" + toString(i) + suffix + ".m", *R);
      }
    }
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Keep(const std::string & ename, const FactoryBase* factory) {
    for (Array<RCP<Level> >::iterator it = Levels_.begin(); it != Levels_.end(); ++it)
      (*it)->Keep(ename, factory);
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Delete(const std::string& ename, const FactoryBase* factory) {
    for (Array<RCP<Level> >::iterator it = Levels_.begin(); it != Levels_.end(); ++it)
      (*it)->Delete(ename, factory);
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::AddKeepFlag(const std::string & ename, const FactoryBase* factory, KeepType keep) {
    for (Array<RCP<Level> >::iterator it = Levels_.begin(); it != Levels_.end(); ++it)
      (*it)->AddKeepFlag(ename, factory, keep);
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::RemoveKeepFlag(const std::string & ename, const FactoryBase* factory, KeepType keep) {
    for (Array<RCP<Level> >::iterator it = Levels_.begin(); it != Levels_.end(); ++it)
      (*it)->RemoveKeepFlag(ename, factory, keep);
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  std::string Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::description() const {
    if ( description_.empty() )
    {
      std::ostringstream out;
      out << BaseClass::description();
      out << "{#levels = " << GetGlobalNumLevels() << ", complexity = " << GetOperatorComplexity() << "}";
      description_ = out.str();
    }
    return description_;
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::describe(Teuchos::FancyOStream& out, const Teuchos::EVerbosityLevel tVerbLevel) const {
    describe(out, toMueLuVerbLevel(tVerbLevel));
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::describe(Teuchos::FancyOStream& out, const VerbLevel verbLevel) const {
    RCP<Operator> A0 = Levels_[0]->template Get<RCP<Operator> >("A");
    RCP<const Teuchos::Comm<int> > comm = A0->getDomainMap()->getComm();

    int numLevels = GetNumLevels();
    RCP<Operator> Ac = Levels_[numLevels-1]->template Get<RCP<Operator> >("A");
    if (Ac.is_null()) {
      // It may happen that we do repartition on the last level, but the matrix
      // is small enough to satisfy "max coarse size" requirement. Then, even
      // though we have the level, the matrix would be null on all but one processors
      numLevels--;
    }
    int root = comm->getRank();

#ifdef HAVE_MPI
    int smartData = numLevels*comm->getSize() + comm->getRank(), maxSmartData;
    reduceAll(*comm, Teuchos::REDUCE_MAX, smartData, Teuchos::ptr(&maxSmartData));
    root = maxSmartData % comm->getSize();
#endif

    // Compute smoother complexity, if needed
    double smoother_comp =-1.0;
    if (verbLevel & (Statistics0 | Test))
      smoother_comp = GetSmootherComplexity();

    std::string outstr;
    if (comm->getRank() == root && verbLevel & (Statistics0 | Test)) {
      std::vector<Xpetra::global_size_t> nnzPerLevel;
      std::vector<Xpetra::global_size_t> rowsPerLevel;
      std::vector<int>                   numProcsPerLevel;
      bool aborted = false;
      for (int i = 0; i < numLevels; i++) {
        TEUCHOS_TEST_FOR_EXCEPTION(!(Levels_[i]->IsAvailable("A")) , Exceptions::RuntimeError,
                                   "Operator A is not available on level " << i);

        RCP<Operator> A = Levels_[i]->template Get<RCP<Operator> >("A");
        TEUCHOS_TEST_FOR_EXCEPTION(A.is_null(), Exceptions::RuntimeError,
                                   "Operator A on level " << i << " is null.");

        RCP<Matrix> Am = rcp_dynamic_cast<Matrix>(A);
        if (Am.is_null()) {
          GetOStream(Warnings0) << "Some level operators are not matrices, statistics calculation aborted" << std::endl;
          aborted = true;
          break;
        }

        Xpetra::global_size_t nnz = Am->getGlobalNumEntries();
        nnzPerLevel     .push_back(nnz);
        rowsPerLevel    .push_back(Am->getGlobalNumRows());
        numProcsPerLevel.push_back(Am->getRowMap()->getComm()->getSize());
      }

      if (!aborted) {
        std::string label = Levels_[0]->getObjectLabel();
        std::ostringstream oss;
        oss << std::setfill(' ');
        oss << "\n--------------------------------------------------------------------------------\n";
        oss << "---                            Multigrid Summary "  << std::setw(28) << std::left << label << "---\n";
        oss << "--------------------------------------------------------------------------------" << std::endl;
        oss << "Number of levels    = " << numLevels << std::endl;
        oss << "Operator complexity = " << std::setprecision(2) << std::setiosflags(std::ios::fixed)
            << GetOperatorComplexity() << std::endl;

        if(smoother_comp!=-1.0) {
          oss << "Smoother complexity = " << std::setprecision(2) << std::setiosflags(std::ios::fixed)
              << smoother_comp << std::endl;
        }

        switch (Cycle_) {
           case VCYCLE:
             oss << "Cycle type          = V" << std::endl;
             break;
           case WCYCLE:
             oss << "Cycle type          = W" << std::endl;
             break;
           default:
             break;
        };
        oss << std::endl;

        Xpetra::global_size_t tt = rowsPerLevel[0];
        int rowspacer = 2; while (tt != 0) { tt /= 10; rowspacer++; }
        tt = nnzPerLevel[0];
        int nnzspacer = 2; while (tt != 0) { tt /= 10; nnzspacer++; }
        tt = numProcsPerLevel[0];
        int npspacer = 2;  while (tt != 0) { tt /= 10; npspacer++; }
        oss  << "level " << std::setw(rowspacer) << " rows " << std::setw(nnzspacer) << " nnz " << " nnz/row" << std::setw(npspacer) << "  c ratio" << "  procs" << std::endl;
        for (size_t i = 0; i < nnzPerLevel.size(); ++i) {
          oss << "  " << i << "  ";
          oss << std::setw(rowspacer) << rowsPerLevel[i];
          oss << std::setw(nnzspacer) << nnzPerLevel[i];
          oss << std::setprecision(2) << std::setiosflags(std::ios::fixed);
          oss << std::setw(9) << as<double>(nnzPerLevel[i]) / rowsPerLevel[i];
          if (i) oss << std::setw(9) << as<double>(rowsPerLevel[i-1])/rowsPerLevel[i];
          else   oss << std::setw(9) << "     ";
          oss << "    " << std::setw(npspacer) << numProcsPerLevel[i] << std::endl;
        }
        oss << std::endl;
        for (int i = 0; i < GetNumLevels(); ++i) {
          RCP<SmootherBase> preSmoo, postSmoo;
          if (Levels_[i]->IsAvailable("PreSmoother"))
            preSmoo = Levels_[i]->template Get< RCP<SmootherBase> >("PreSmoother");
          if (Levels_[i]->IsAvailable("PostSmoother"))
            postSmoo = Levels_[i]->template Get< RCP<SmootherBase> >("PostSmoother");

          if (preSmoo != null && preSmoo == postSmoo)
            oss << "Smoother (level " << i << ") both : " << preSmoo->description() << std::endl;
          else {
            oss << "Smoother (level " << i << ") pre  : "
                << (preSmoo != null ?  preSmoo->description() : "no smoother") << std::endl;
            oss << "Smoother (level " << i << ") post : "
                << (postSmoo != null ?  postSmoo->description() : "no smoother") << std::endl;
          }

          oss << std::endl;
        }

        outstr = oss.str();
      }
    }

#ifdef HAVE_MPI
    RCP<const Teuchos::MpiComm<int> > mpiComm = rcp_dynamic_cast<const Teuchos::MpiComm<int> >(comm);
    MPI_Comm rawComm = (*mpiComm->getRawMpiComm())();

    int strLength = outstr.size();
    MPI_Bcast(&strLength, 1, MPI_INT, root, rawComm);
    if (comm->getRank() != root)
      outstr.resize(strLength);
    MPI_Bcast(&outstr[0], strLength, MPI_CHAR, root, rawComm);
#endif

    out << outstr;
  }

  // NOTE: at some point this should be replaced by a friend operator <<
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::print(std::ostream& out, const VerbLevel verbLevel) const {
    Teuchos::OSTab tab2(out);
    for (int i = 0; i < GetNumLevels(); ++i)
      Levels_[i]->print(out, verbLevel);
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::IsPreconditioner(const bool flag) {
    isPreconditioner_ = flag;
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::DumpCurrentGraph() const {
    if (GetProcRankVerbose() != 0)
      return;
#if defined(HAVE_MUELU_BOOST) && defined(HAVE_MUELU_BOOST_FOR_REAL) && defined(BOOST_VERSION) && (BOOST_VERSION >= 104400)
    BoostGraph      graph;

    BoostProperties dp;
    dp.property("label", boost::get(boost::vertex_name,  graph));
    dp.property("id",    boost::get(boost::vertex_index, graph));
    dp.property("label", boost::get(boost::edge_name,    graph));
    dp.property("color", boost::get(boost::edge_color,   graph));

    // create local maps
    std::map<const FactoryBase*, BoostVertex>                                     vindices;
    typedef std::map<std::pair<BoostVertex,BoostVertex>, std::string> emap; emap  edges;

    static int call_id=0;

    RCP<Operator> A = Levels_[0]->template Get<RCP<Operator> >("A");
    int rank = A->getDomainMap()->getComm()->getRank();

    //    printf("[%d] CMS: ----------------------\n",rank);
    for (int i = dumpLevel_; i <= dumpLevel_+1 && i < GetNumLevels(); i++) {
      edges.clear();
      Levels_[i]->UpdateGraph(vindices, edges, dp, graph);

      for (emap::const_iterator eit = edges.begin(); eit != edges.end(); eit++) {
        std::pair<BoostEdge, bool> boost_edge = boost::add_edge(eit->first.first, eit->first.second, graph);
  //  printf("[%d] CMS:   Hierarchy, adding edge (%d->%d) %d\n",rank,(int)eit->first.first,(int)eit->first.second,(int)boost_edge.second);
  // Because xdot.py views 'Graph' as a keyword
  if(eit->second==std::string("Graph")) boost::put("label", dp, boost_edge.first, std::string("Graph_"));
        else boost::put("label", dp, boost_edge.first, eit->second);
        if (i == dumpLevel_)
          boost::put("color", dp, boost_edge.first, std::string("red"));
        else
          boost::put("color", dp, boost_edge.first, std::string("blue"));
      }
    }

#if 0
    std::ostringstream legend;
    legend << "< <TABLE BORDER=\"0\" CELLBORDER=\"1\" CELLSPACING=\"0\" CELLPADDING=\"4\"> \
               <TR><TD COLSPAN=\"2\">Legend</TD></TR> \
               <TR><TD><FONT color=\"red\">Level " << dumpLevel_ << "</FONT></TD><TD><FONT color=\"blue\">Level " << dumpLevel_+1 << "</FONT></TD></TR> \
               </TABLE> >";
    BoostVertex boost_vertex = boost::add_vertex(graph);
    boost::put("label", dp, boost_vertex, legend.str());
#endif

    std::ofstream out(dumpFile_.c_str() +std::to_string(call_id)+std::string("_")+ std::to_string(rank) + std::string(".dot"));
    boost::write_graphviz_dp(out, graph, dp, std::string("id"));
    out.close();
    call_id++;
#else
    GetOStream(Errors) <<  "Dependency graph output requires boost and MueLu_ENABLE_Boost_for_real" << std::endl;
#endif
  }

  // Enforce that coordinate vector's map is consistent with that of A
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::ReplaceCoordinateMap(Level& level) {
    RCP<Operator> Ao = level.Get<RCP<Operator> >("A");
    RCP<Matrix>   A  = rcp_dynamic_cast<Matrix>(Ao);
    if (A.is_null()) {
      GetOStream(Warnings1) << "Hierarchy::ReplaceCoordinateMap: operator is not a matrix, skipping..." << std::endl;
      return;
    }
    if(Teuchos::rcp_dynamic_cast<BlockedCrsMatrix>(A) != Teuchos::null) {
      GetOStream(Warnings1) << "Hierarchy::ReplaceCoordinateMap: operator is a BlockedCrsMatrix, skipping..." << std::endl;
      return;
    }

    typedef Xpetra::MultiVector<typename Teuchos::ScalarTraits<Scalar>::coordinateType,LO,GO,NO> xdMV;

    RCP<xdMV> coords = level.Get<RCP<xdMV> >("Coordinates");

    if (A->getRowMap()->isSameAs(*(coords->getMap()))) {
      GetOStream(Warnings1) << "Hierarchy::ReplaceCoordinateMap: matrix and coordinates maps are same, skipping..." << std::endl;
      return;
    }

    if (A->IsView("stridedMaps") && rcp_dynamic_cast<const StridedMap>(A->getRowMap("stridedMaps")) != Teuchos::null) {
      RCP<const StridedMap> stridedRowMap = rcp_dynamic_cast<const StridedMap>(A->getRowMap("stridedMaps"));

      // It is better to through an exceptions if maps may be inconsistent, than to ignore it and experience unfathomable breakdowns
      TEUCHOS_TEST_FOR_EXCEPTION(stridedRowMap->getStridedBlockId() != -1 || stridedRowMap->getOffset() != 0,
                                 Exceptions::RuntimeError, "Hierarchy::ReplaceCoordinateMap: nontrivial maps (block id = " << stridedRowMap->getStridedBlockId()
                                 << ", offset = " << stridedRowMap->getOffset() << ")");
    }

    GetOStream(Runtime1) << "Replacing coordinate map" << std::endl;

    size_t blkSize = A->GetFixedBlockSize();

    RCP<const Map> nodeMap = A->getRowMap();
    if (blkSize > 1) {
      // Create a nodal map, as coordinates have not been expanded to a DOF map yet.
      RCP<const Map> dofMap       = A->getRowMap();
      GO             indexBase    = dofMap->getIndexBase();
      size_t         numLocalDOFs = dofMap->getNodeNumElements();
      TEUCHOS_TEST_FOR_EXCEPTION(numLocalDOFs % blkSize, Exceptions::RuntimeError,
        "Hierarchy::ReplaceCoordinateMap: block size (" << blkSize << ") is incompatible with the number of local dofs in a row map (" << numLocalDOFs);
      ArrayView<const GO> GIDs = dofMap->getNodeElementList();

      Array<GO> nodeGIDs(numLocalDOFs/blkSize);
      for (size_t i = 0; i < numLocalDOFs; i += blkSize)
        nodeGIDs[i/blkSize] = (GIDs[i] - indexBase)/blkSize + indexBase;

      Xpetra::global_size_t INVALID = Teuchos::OrdinalTraits<Xpetra::global_size_t>::invalid();
      nodeMap = MapFactory::Build(dofMap->lib(), INVALID, nodeGIDs(), indexBase, dofMap->getComm());
    } else {
      // blkSize == 1
      // Check whether the length of vectors fits to the size of A
      // If yes, make sure that the maps are matching
      // If no, throw a warning but do not touch the Coordinates
      if(coords->getLocalLength() != A->getRowMap()->getNodeNumElements()) {
        GetOStream(Warnings) << "Coordinate vector does not match row map of matrix A!" << std::endl;
        return;
      }
    }

    Array<ArrayView<const typename Teuchos::ScalarTraits<Scalar>::coordinateType> >      coordDataView;
    std::vector<ArrayRCP<const typename Teuchos::ScalarTraits<Scalar>::coordinateType> > coordData;
    for (size_t i = 0; i < coords->getNumVectors(); i++) {
      coordData.push_back(coords->getData(i));
      coordDataView.push_back(coordData[i]());
    }

    RCP<xdMV> newCoords = Xpetra::MultiVectorFactory<typename Teuchos::ScalarTraits<Scalar>::coordinateType,LO,GO,NO>::Build(nodeMap, coordDataView(), coords->getNumVectors());
    level.Set("Coordinates", newCoords);
  }

  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::AllocateLevelMultiVectors(int numvecs) {
    int N = Levels_.size();  
    if( (sizeOfAllocatedLevelMultiVectors_ == numvecs && residual_.size() == N) || numvecs<=0 ) return;

    // If, somehow, we changed the number of levels, delete everything first
    if(residual_.size() != N) {
      DeleteLevelMultiVectors();

      residual_.resize(N);
      coarseRhs_.resize(N);
      coarseX_.resize(N);
      coarseImport_.resize(N);
      coarseExport_.resize(N);
      correction_.resize(N);
    }

    for(int i=0; i<N; i++) {
      RCP<Operator> A = Levels_[i]->template Get< RCP<Operator> >("A");
      if(!A.is_null()) {
        // This dance is because we allow A to have a BlockedMap and X/B to have (compatible) non-blocked map
        RCP<const BlockedCrsMatrix> A_as_blocked = Teuchos::rcp_dynamic_cast<const BlockedCrsMatrix>(A);
        RCP<const Map> Arm = A->getRangeMap();
        RCP<const Map> Adm = A->getDomainMap();
        if(!A_as_blocked.is_null()) { 
          Adm = A_as_blocked->getFullDomainMap();
        }

        // This is zero'd by default since it is filled via an operator apply        
        residual_[i] = MultiVectorFactory::Build(Arm, numvecs, true);
        correction_[i] = MultiVectorFactory::Build(Adm, numvecs, false);
      }

      if(i+1<N) {
        // This is zero'd by default since it is filled via an operator apply
        if(implicitTranspose_) {
          RCP<Operator> P = Levels_[i+1]->template Get< RCP<Operator> >("P");
          if(!P.is_null()) coarseRhs_[i] = MultiVectorFactory::Build(P->getDomainMap(),numvecs,true);
        } else {
          RCP<Operator> R = Levels_[i+1]->template Get< RCP<Operator> >("R");
          if(!R.is_null()) coarseRhs_[i] = MultiVectorFactory::Build(R->getRangeMap(),numvecs,true);
        }


        RCP<const Import> importer;
        if(Levels_[i+1]->IsAvailable("Importer"))
          importer = Levels_[i+1]->template Get< RCP<const Import> >("Importer");
        if (doPRrebalance_ || importer.is_null()) 
          coarseX_[i] = MultiVectorFactory::Build(coarseRhs_[i]->getMap(),numvecs,false);
        else {
          coarseImport_[i] = MultiVectorFactory::Build(importer->getTargetMap(), numvecs,false);
          coarseExport_[i] = MultiVectorFactory::Build(importer->getSourceMap(), numvecs,false);
          coarseX_[i] = MultiVectorFactory::Build(importer->getTargetMap(),numvecs,false);
        }
      }
    }
    sizeOfAllocatedLevelMultiVectors_ = numvecs;
  }


template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
void Hierarchy<Scalar, LocalOrdinal, GlobalOrdinal, Node>::DeleteLevelMultiVectors() {
  if(sizeOfAllocatedLevelMultiVectors_==0) return;
  residual_.resize(0);
  coarseRhs_.resize(0);
  coarseX_.resize(0);
  coarseImport_.resize(0);
  coarseExport_.resize(0);
  correction_.resize(0);
  sizeOfAllocatedLevelMultiVectors_ = 0;
}


} //namespace MueLu

#endif // MUELU_HIERARCHY_DEF_HPP